Research map / Probabilistic and scientific modeling / Differential equations and PDEs
"Hey, that’s not an ODE": Faster ODE Adjoints via Seminorms
Patrick Kidger, Ricky T. Q. Chen, Terry J Lyons
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the adjoint seminorm instead of the default norm | Neural CDE · Speech Commands | backward NFE: 14.36 ±1.09 -> 8.67 ±1.60; 30.67 ±2.48 -> 12.75 ±2.00; 77.95 ±4.47 -> 29.39 ±0.80 | Improved |
| Use the adjoint seminorm instead of the default norm | Neural CDE · Speech Commands | test accuracy: 92.6 ±0.4 -> 92.5 ±0.5; 92.8 ±0.4 -> 92.5 ±0.5; 92.4 ±0.7 -> 92.9 ±0.4 | No clear effect |
| Use the adjoint seminorm instead of the default norm | Continuous Normalizing Flow · CIFAR-10 (d_h=64) | backward NFE: 41.65 ±1.97 -> 38.93 ±1.32 | Improved |
| Use the adjoint seminorm instead of the default norm | Continuous Normalizing Flow · MNIST | backward NFE: 41.50 ±2.35 -> 37.03 ±0.96; 44.24 ±1.78 -> 37.85 ±0.94; 48.60 ±2.30 -> 41.84 ±1.92 | Improved |
| Use the adjoint seminorm instead of the default norm | Continuous Normalizing Flow · CIFAR-10 and MNIST | test bits/dim: 3.3492 ±0.0059 -> 3.3428 ±0.0121; 0.9968 ±0.0020 -> 0.9942 ±0.0013; 0.9623 ±0.0020 -> 0.9603 ±0.0021; 0.9439 ±0.0030 -> | Mixed |
| Use the adjoint seminorm instead of the default norm | Symplectic ODE-Net · Fully-actuated double pendulum | backward NFE: 46.45 ±0.01 -> 26.55 ±0.01 | Improved |
| Use the adjoint seminorm instead of the default norm | Symplectic ODE-Net · Fully-actuated double pendulum | test L2 loss: 1.247 ±0.520 -> 2.995 ±2.190 | Hurt |
| Exclude a_theta and a_t channels from step-size control | Neural differential equation adjoint solve · Neural CDE, continuous normalizing flows, and Hamiltonian dynamics | backward function evaluations: 40% fewer function evaluations (median); as much as 62% fewer function evaluations | Improved |
Closest papers by meaning
- Do Residual Neural Networks discretize Neural Ordinary Differential Equations? (NeurIPS 2022 · similarity 0.89)
- Efficient Training of Neural Fractional-Order Differential Equation via Adjoint Backpropagation (AAAI 2025 · similarity 0.88)
- How Deep Do We Need: Accelerating Training and Inference of Neural ODEs via Control Perspective (ICML 2024 · similarity 0.88)
- Global Convergence in Neural ODEs: Impact of Activation Functions (ICLR 2025 · similarity 0.88)
- AdjointDEIS: Efficient Gradients for Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Breakeven complexity: A new perspective on neural partial differential equation solvers (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
